igie

DINOv2 ViT-S/14 (IGIE)

Model Description

DINOv2 is a self-supervised vision Transformer model designed for general-purpose image representation learning. This example uses the ViT-S/14 pretrained backbone for ImageNet linear evaluation: the frozen backbone extracts intermediate features, and a linear classifier is trained and evaluated on top of them.

Supported Environments

GPU	IXUCA SDK	Release
MR-V100	4.4.0	26.06

Model Preparation

visit https://github.com/facebookresearch/dinov2 to see the official repo.

Prepare Resources

Pretrained model: https://dl.fbaipublicfiles.com/dinov2/dinov2_vits14/dinov2_vits14_pretrain.pth

Dataset: https://www.image-net.org/download.php to download the ImageNet-1K dataset.

Install Dependencies

git clone https://github.com/facebookresearch/dinov2.git
cp -r eval dinov2/dinov2
cp dinov2-patch/* dinov2/
pip3 install omegaconf==2.3.0 fvcore iopath submitit==1.5.4 torchmetrics

Model Conversion

cd dinov2

wget https://github.com/facebookresearch/dinov2/files/13427966/labels.txt -O /mnt/deepspark/data/datasets/imagenet
python3 prepare_datasets.py

python3 export.py \
    --config-file dinov2/configs/eval/vits14_pretrain.yaml \
    --pretrained-weights dinov2_vits14_pretrain.pth \
    --onnx-path dinov2_vits14_pretrain.onnx \
    --input-size 224 \
    --n-last-blocks 4 \
    --device cuda

export batchsize=64
python3 build_engine.py                     \
    --model_path dinov2_vits14_pretrain.onnx              \
    --input input:${batchsize},3,224,224    \
    --precision fp16                        \
    --engine_path dinov2_vits14_pretrain_bs_${batchsize}_fp16.so

Model Inference

export PYTHONPATH=/path/to/dinov2:${PYTHONPATH}
export IMAGENET_1K=/path/to/imagenet

# ACC
bash scripts/infer_dinov2_fp16_accuracy.sh

# Performence
bash scripts/infer_dinov2_fp16_performance.sh

Model Results

Model	Task	BatchSize	Precision	FPS	Accuracy
DINOv2 ViT-S/14	Linear Eval	64	FP16	889.542	81.1%

References

• dinov2